Scientific AI agents need to preserve the decisions between experiments

An automated laboratory can read a paper, propose a formulation, run a test and choose the next experiment. The difficult question is what survives that process as a scientific record. A new primer in Nature Reviews Methods Primers places the decisions between experiments at the centre of agentic research: a result becomes more valuable when another group can trace how it was reached.[1]

Published on 22 October, the primer describes systems that combine task planning, memory, tool use and communication. They can connect literature search, physics-based simulations and automated experimental platforms. Its contribution is a framework for organising those activities, including human-supervised and more autonomous configurations, rather than a new experimental result from a single AI laboratory.[1]

The hidden middle of a discovery

A conventional research article often presents a clean sequence: hypothesis, method, result, interpretation. Real work is messier. Scientists choose among competing techniques, discard failed preparations, correct instruments, revise thresholds and sometimes decide that a signal is too ambiguous to trust. When an AI agent performs some of those steps, it can create a polished narrative that omits the branching path. A result without its intermediate decisions may be very difficult to evaluate or reproduce.[1]

A traceable agentic workflow should therefore preserve the exact inputs, software versions, prompts or policies, retrieved documents, tool commands, instrument settings, raw measurements and decisions that altered the next action. This is not bureaucratic decoration. A changed temperature setting might explain why a catalyst appeared to work. An incorrect database entry might have directed the model to the wrong compound family. A researcher cannot judge the science if crucial actions are hidden beneath the sentence 'the agent investigated the candidates.'[1]

Who can act, and who can approve

The review distinguishes human-in-the-loop systems from configurations designed for greater autonomy. In a tightly controlled workflow, the agent may search papers and propose experiments while an expert approves expensive or hazardous operations. In a more automated setup, the agent can submit jobs to a simulation server or experimental platform and use the results to select another test. The degree of autonomy depends as much on permitted actions and oversight as on the capabilities of the underlying language model.[1]

Specialization can help: one component might retrieve relevant research, another translate a candidate into an instrument recipe, another audit numerical consistency. But adding agents also creates communication and accountability problems. If two agents agree because they share the same misconception, consensus is not independent verification. If one tool returns a corrupted measurement and the reasoning agent treats it as authoritative, the entire workflow may be confidently misdirected. Human review needs to focus on consequential steps rather than simply count how many assistants participated.[1]

Keeping prediction and measurement connected

A model can predict how a molecule might behave using a learned representation of existing data. It can run a simulation that approximates physical behavior. Neither is the same as producing a new sample and observing it under conditions another laboratory can recreate. Closed-loop systems become scientifically meaningful when measurement feeds back into design and when the final claims survive independent checks. They can be powerful ways to decide which experiment to attempt next, without turning theoretical suggestions into measured discoveries by phrasing alone.[1]

The primer identifies hallucination, tool misuse, limited interpretability and safety as bottlenecks. In a laboratory, these failures propagate through actions: an invented premise can determine a candidate list, and an incorrect instrument command can invalidate the measurement used to select the next test. Recording that chain makes the failure diagnosable.[1]

A method another laboratory can use

A useful challenge for any self-described AI scientist is simple to state: can a separate group follow the complete record, rerun the method and determine whether the conclusion holds? That group should be able to distinguish automated decisions from human interventions, predictions from observations, and abandoned attempts from successful ones. If not, the system may have automated activity without strengthening evidence.[1]

A well-documented agentic workflow could become a reusable research method: a sequence of retrieval, computation and measurement that another laboratory can examine and adapt. That is a more durable contribution than a polished AI-written abstract, because it preserves the route from a question to an answer.[1]

Report an error or suggest a correction